Service generation-oriented network scheduling method and system

Abstract

The invention provides a service generation-oriented network scheduling method and system, and the method comprises the steps: receiving a service request task, and analyzing the service request task to obtain task information which comprises a request node and a resource demand of the task; constructing a resource competition graph based on the resource demand of each service request task, taking each service request task as a node in the resource competition graph, and if the resource demands of any two service request tasks conflict, constructing an edge; constructing a Prompt sequence based on all task information and the resource competition atlas, inputting the Prompt sequence into a preset large decision model, and outputting an initial strategy; constructing a prediction vector based on a task request time point in the task information, and outputting a task prediction result by adopting a time sequence prediction model; inputting the resource competition atlas into a preset graph neural network model, and outputting a graph embedded feature vector; and outputting a final strategy by adopting a reinforcement learning algorithm based on the initial strategy, the task prediction result and the graph embedding feature vector.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of network service, and in particular to a network scheduling method and system for service generation. BACKGROUND

[0002] In recent years, with the rapid development of large language model (LLM, Large Language Model) and multi-modal intelligent agent, etc. generative artificial intelligence (GAI, Generative Artificial Intelligence) technology, network systems are facing unprecedented demand and challenges in infrastructure and service capabilities. On the one hand, such emerging technologies usually have the characteristics of large parameter size and complex reasoning chain, which puts forward the demand for large-scale heterogeneous computing power, low-latency high-throughput network and adaptive fast storage, etc. multi-dimensional resources for the underlying infrastructure. On the other hand, different business types (such as autonomous driving, real-time video generation, personalized large model service, etc.) present differentiated service requirements, and there are significant differences in delay, security, reliability guarantee, etc. These trends collectively push the network system to accelerate from the traditional data transmission direction to a new stage of resource coordination and service awareness.

[0003] The traditional cloud computing and edge computing paradigm is difficult to balance the flexibility of large-scale computing scheduling and diversified service quality requirements under the constraints of centralized architecture and limited coordination range. SUMMARY

[0004] In view of this, the embodiments of the present application provide a network scheduling method and system for service generation to eliminate or improve one or more defects in the prior art.

[0005] One aspect of the present application provides a network scheduling method for service generation, the method is based on a service network comprising a plurality of processing nodes, the steps of the method comprising: receiving a service request task and parsing to obtain task information, the task information including a request node and resource requirements of the task; constructing a resource competition graph based on the resource requirements of each service request task, in the resource competition graph, each service request task is regarded as a node, and if the resource requirements of any two service request tasks conflict, an edge is constructed between the two nodes; constructing a Prompt sequence based on all the task information and the resource competition graph, inputting the Prompt sequence into a preset large decision model, and the large decision model outputs an initial strategy; constructing a prediction vector based on the task request time point in the task information, and outputting a task prediction result using a time series prediction model; input the resource competition graph into a preset graph neural network model, and the graph neural network model outputs a graph embedding feature vector; Based on the initial strategy, the task prediction result and the graph embedding feature vector, a reinforcement learning algorithm is used to output a final strategy.

[0006] With the above scheme, the present scheme takes a large decision model and a reinforcement learning algorithm as the core, responsible for generating and optimizing network service strategies. First, the unified coding of the fused business demand and resource competition graph is performed, and the Prompt sequence is constructed combining the resource state, constraint condition and historical strategy experience and other multi-modal information to drive the LDM to generate a structured strategy draft covering task deployment, resource allocation and path selection, and to perform feasibility verification. Subsequently, a time series prediction model (such as Informer, LSTM) is introduced to predict the change trend of the network resource load, and the graph neural network (GNN, Graph Neural Network) is used to extract the task coupling features and conflict paths from the resource competition graph. Finally, the system inputs the strategy draft as the initial strategy of the reinforcement learning agent, combines the prediction result and the graph structure feature embedding state space, and carries out strategy training and evolution through the "state-action-reward" mechanism to generate a candidate strategy set that can adapt to different performance requirements and resource constraints, and finally selects the strategy, which takes into account the flexibility of large-scale computing power scheduling and diversified service quality requirements.

[0007] In some embodiments of the present application, if the resource requirements of any two service request tasks conflict, in the step of constructing edges between two nodes, the resource types required by the service request tasks are read from the task information, and it is determined whether the resource types required by any two service request tasks overlap. If there is overlap, it is determined that the resource requirements of the two service request tasks conflict.

[0008] In some embodiments of the present application, if the resource requirements of any two service request tasks conflict, the step of constructing edges between two nodes further includes constructing a label including identification information on the edge between the two nodes, the identification information including the node in the service network requested by the service request task, the remaining resource quantity in the node in the service network requested by the service request task, the resource quantity requested by the service request task to the node in the service network, and the competing resource quantity.

[0009] In some embodiments of the present application, in the step of constructing a Prompt sequence based on all the task information and resource competition graph, a prompt word engineering technology is used to construct a Prompt sequence based on all the task information and resource competition graph.

[0010] In some embodiments of the present application, in the step of outputting the final strategy by the reinforcement learning algorithm based on the initial strategy, the task prediction result and the graph embedding feature vector, the initial action of the reinforcement learning algorithm is constructed based on the initial strategy, and the initial state of the reinforcement learning algorithm is constructed based on the task prediction result and the graph embedding feature vector.

[0011] In some embodiments of the present application, in the step of outputting the final strategy by the reinforcement learning algorithm, the reinforcement learning algorithm outputs a strategy in the processing of each round; In the processing of each round, the updated state is determined based on the initial state of each round and the selected action, the business target achievement index, the cost item, the energy consumption item, the constraint penalty and the stability item are determined based on the updated state, and the reward value is determined based on the business target achievement index, the cost item, the energy consumption item, the constraint penalty and the stability.

[0012] In some embodiments of the present application, in the step of determining the reward value based on the business target achievement index, the cost item, the energy consumption item, the constraint penalty and the stability, the following formula is used for calculation: Reward value = α * S - β * C - γ * E - λ * V - μ * M; Wherein, S represents the business target achievement index, α represents the weight corresponding to the business target achievement index, C represents the cost item, β represents the weight corresponding to the cost item, E represents the energy consumption item, γ represents the weight corresponding to the energy consumption item, V represents the constraint penalty, λ represents the weight corresponding to the constraint penalty, M represents the stability item, and μ represents the weight corresponding to the stability item.

[0013] In some embodiments of the present application, in the step of outputting the final strategy by the reinforcement learning algorithm, the strategies output by the reinforcement learning algorithm in the processing of all rounds are collected as an initial strategy draft, the business target achievement indexes of each strategy are sorted, and the final strategy is selected based on the sorting result.

[0014] In some embodiments of the present application, in the step of receiving the service request task and parsing to obtain the task information, the service request task is parsed by a large language model to obtain the task information; The method further comprises the step of verifying the final strategy by a twin mapping module in a twin environment. The method further comprises the step of configuring the network according to the final strategy by a service orchestration module.

[0015] A second aspect of the present invention also provides a service-oriented network scheduling system, the system comprising a computer device including a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions stored in the memory, and the system implementing the steps of the method described above when the computer instructions are executed by the processor.

[0016] A third aspect of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned service-oriented network scheduling method.

[0017] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the text, or may be learned by practice of the invention. The objects and other advantages of the invention will become apparent from the description and the accompanying drawings.

[0018] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0019] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to limit the scope of the invention.

[0020] Figure 1 This is a schematic diagram illustrating one implementation of the service-oriented network scheduling method of this solution; Figure 2 This is a schematic diagram of an existing technology network; Figure 3 This is a schematic diagram of the overall architecture of this solution; Figure 4 This is a schematic diagram of the processing logic of this solution; Figure 5 This is a schematic diagram of the processing framework of this solution; Figure 6 This is a schematic diagram of the core processing steps of this solution; Figure 7 This is a schematic diagram of the model processing framework for this solution. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0022] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0023] Introduction to existing technologies: Traditional cloud computing and edge computing paradigms, constrained by centralized architecture and limited collaborative scope, struggle to flexibly meet both large-scale computing power scheduling and diverse service quality requirements. Therefore, as... Figure 2 As shown, computing-network convergence technologies and service-customized architectures, typically represented by Computing Power Network (CPN) and Service-Customized Networking (SCN), have become research hotspots.

[0024] Computing-network integration is a key practical path for the concept of computing-network convergence. It aims to achieve on-demand accessibility, dynamic scheduling, and collaborative orchestration of computing power across cloud, edge, and endpoint devices through deep collaboration between computing and network resources. Its core is to overcome the limitations of the separation of communication and computing in traditional networks, building service-oriented global collaborative capabilities to support efficient resource provisioning and agile response in complex business scenarios. Currently, major domestic operators and telecommunications companies have conducted extensive research in areas such as network slicing, computing power deployment, network function virtualization, and computing resource scheduling, striving to achieve dynamic collaboration between network and computing resources and promoting the transformation of this technology from the proof-of-concept stage to actual engineering deployment. Simultaneously, academia has conducted extensive research in areas such as architecture design and resource management, aiming to drive computing-network convergence towards a more automated and precise intelligent direction.

[0025] Customized service networks (CSUs) represent another new type of network architecture geared towards service capability evolution. Their core idea is to dynamically customize network functions and resource paths based on user service needs, enabling flexible assurance of differentiated services. Compared to traditional resource-centric network construction, CSUs emphasize the "network as a service" concept, dynamically adapting to multi-source service requests through the orchestration of network functional modules to achieve flexible construction and fine-grained delivery of service capabilities. The development of CSUs benefits from the mature application of technologies such as network function virtualization and software-defined networking. Early research focused on decoupling network functions and abstracting resources, providing support for the programmability and flexibility of infrastructure.

[0026] While existing research has laid the theoretical and practical foundation for computing power networks, they still face numerous challenges in large-scale deployment. On the one hand, computing resources are highly heterogeneous, with significant differences in capabilities between nodes, making unified abstraction and collaborative management difficult. On the other hand, traditional orchestration and scheduling mechanisms suffer from response lag and policy rigidity, making it difficult to meet the service assurance needs of highly dynamic business scenarios. Furthermore, computing power networks still lack systematic support in areas such as resource awareness, intelligent decision-making, and autonomous control. Therefore, how to build intelligent computing power networks with self-awareness, self-learning, and self-optimization capabilities has become a key technological direction for promoting the deep integration of computing and networks.

[0027] Service-customized networks also face several limitations in practical deployments. First, current service orchestration largely relies on predefined models and static strategies, lacking the ability to adapt to real-time business changes. Second, orchestration logic is primarily based on structured rules, lacking flexibility and intelligence, making it difficult to support complex service scenarios involving multiple modalities, objectives, and spatiotemporal scales. Furthermore, in terms of assurance mechanisms, support for resource conflicts, service preemption, and dynamic demand adjustments remains relatively weak. Therefore, driving the evolution of service-customized networks towards intelligence and automation urgently requires achieving deep understanding of service semantics, dynamic optimization of resource decisions, and closed-loop management of the service lifecycle, thereby providing more efficient and reliable network support capabilities for diverse services.

[0028] like Figure 7 As shown, closed-loop optimization consists of three sub-components: optimization triggering, strategy determination, and optimization execution. 1. Optimization Trigger. This component is responsible for real-time monitoring of the model's running status and triggering the optimization process when potential performance degradation risks are detected, ensuring the model's continued effectiveness and service stability. The system primarily senses three types of signals: First, input distribution drift signals from the intent parsing module, used to identify input semantic shifts caused by changes in task type, adjustments in user expression, or resource structure restructuring; second, model structure aging signals from the policy generation module, reflecting performance degradation trends caused by parameter saturation, decreased generalization ability, or increased multi-task interference; and third, policy performance degradation signals provided by the twin mapping module, used to detect abnormal behaviors such as service execution failures and increased prediction bias. The system continuously senses and dynamically judges the above signals based on preset thresholds to determine whether to enter the model optimization phase.

[0029] 2. Strategy Determination. Upon triggering the model optimization process, the system enters the strategy determination phase to decide whether to employ a fine-tuning mechanism or a transfer learning mechanism. First, the system summarizes current business input and resource status information, assesses the degree of task disturbance and resource structure changes, and determines whether there are significant input shifts or topology reconstructions. Subsequently, it analyzes the model parameter update frequency, gradient change trends, and correlation indicators to determine if there are capacity bottlenecks or a decline in expressive power. Simultaneously, it compares the execution effects of existing strategies in real deployment environments and twin simulation environments, focusing on key performance indicators such as task completion rate and resource utilization to quantify the degree of performance degradation. Finally, the system uses a rule engine to perform weighted scoring based on the impact of various indicators on service quality. When the overall score is below a threshold, it indicates that the model has only experienced mild degradation, and the system will execute a rapid fine-tuning strategy to achieve low-overhead, high-efficiency adaptation. When the score is above the threshold, it indicates that the model's capabilities can no longer support the current business complexity, and the system will initiate a transfer learning mechanism for structural reconstruction and capability rebuilding.

[0030] 3. Optimized Execution. Based on the policy determination results, the system will initiate the corresponding optimization execution process. If a rapid fine-tuning mechanism is adopted, the system will load the current base version of the model, freeze most parameters, and only perform lightweight updates to the policy output layer or key sub-modules. Training data comes from recently collected samples and feedback information, and the training method uses efficient parameter fine-tuning techniques such as LoRA and QLoRA to reduce resource consumption and shorten optimization latency. If a transfer learning mechanism is adopted, the system will comprehensively apply domain adaptation and meta-learning techniques to reconstruct the model structure and key modules, completing capability transfer and representation reconstruction. Training data covers both original knowledge and new scenario data to improve the model's generalization ability in complex environments. After optimization, the system will verify the model performance in a twin environment to ensure that it meets deployment requirements in terms of stability and controllability, ultimately achieving gradual deployment of the policy and ensuring service continuity and system security.

[0031] In summary, through the dynamic synergy of efficient model fine-tuning and adaptive migration mechanisms, rapid response to environmental changes and continuous evolution of policy capabilities can be achieved without compromising the model's fundamental capabilities. This technology is not only a key support for intelligent service generation but also the core engine enabling this solution to possess environmental awareness, risk response, and continuous optimization capabilities, holding significant strategic importance for realizing an efficient, stable, and adaptive next-generation intelligent network architecture.

[0032] Meanwhile, with the rapid development of GAI technology, by learning the distribution characteristics of large-scale data, it can autonomously generate multimodal content such as text, images, and speech based on understanding and processing input information, demonstrating creativity, generalization, and expressive ability far exceeding traditional discriminative models. Therefore, deeply embedding GAI technology into network architecture can drive an integrated intelligent closed loop of "perception-reasoning-decision-optimization," such as... Figure 5 As shown, this approach achieves a synergistic enhancement of multi-dimensional capabilities such as intent parsing, policy generation, anomaly detection, and system adaptation, thereby significantly improving the network's adaptive orchestration and optimization capabilities, effectively supporting efficient service delivery and differentiated service assurance. Furthermore, by incorporating technologies such as LLM, reinforcement learning (RL), and digital twin (DT), it can further enrich the cognitive reasoning and dynamic optimization capabilities of intelligent networks, providing a core driving force for next-generation network service architectures.

[0033] To this end, this invention proposes a future-oriented intelligent network architecture—a service-generation network—based on GAI technology and the concept of computing-network convergence. With intelligent generation of network services as its core, it aims to achieve intelligent control over the entire process of business intent parsing, service policy generation, and network configuration execution. Through this design, this invention not only provides intelligent decision-making and adaptive optimization capabilities for computing networks but also compensates for the shortcomings of service-customized networks in intelligent semantic parsing and dynamic orchestration, laying the technical foundation for a future efficient, reliable, and differentiated network service system.

[0034] like Figure 1 , 3 As shown in Figure 6, this invention proposes a service-oriented network scheduling method. The method is based on a service network comprising multiple processing nodes, and the steps of the method include: Step S100: Receive a service request task and parse it to obtain task information, the task information including the request node and the resource requirements of the task; Specifically, this includes the semantic alignment step, accessing the semantic ontology library, mapping "low latency" to <30ms, and "high reliability" to packet loss rate <0.01%.

[0035] Normalize synonyms (such as "low latency" and "latency sensitive").

[0036] Output: Structured resource demand vector.

[0037] Step S200: Construct a resource contention graph based on the resource requirements of each service request task. In the resource contention graph, each service request task is treated as a node. If the resource requirements of any two service request tasks conflict, an edge is constructed between the two nodes. Specifically, in the process of constructing the resource competition map, a unified demand set and resource status are achieved.

[0038] Operational details: Each task is modeled as a node in a graph. When multiple tasks contend for the same resource (such as an edge node GPU), edges are established to represent the conflict relationship. Conflict intensity is calculated as: conflict = Σrequest - available. Edge attributes are: {resource_id: Edge-GPU-A, available:3, request:4, conflict:1}. Output: Task-to-task conflict intensity graph. Output: Resource contention graph (with node and edge attributes).

[0039] For example, user T1 requests GPU: 2 (latency < 30ms), T2 requests GPU: 2 (latency < 50ms), T3 requests GPU: 1 (latency < 100ms); the remaining GPUs on the edge node are 3.

[0040] process: Standardize the format → Use JSON.

[0041] Semantic alignment → “Low latency” is mapped to <30ms.

[0042] Conflict modeling → Construct a conflict map and label the GPU conflict intensity as 2.

[0043] Output: Conflict graph, containing three tasks T1, T2, and T3 and their competition with the GPU.

[0044] Using the above approach, the construction of the resource contention graph is a key step before strategy generation. It aims to identify and quantify potential resource contention relationships in advance based on user task requirements and the current system resource status.

[0045] First, during the demand aggregation phase, the system has already obtained the structured task requirements submitted by the user (such as the number of GPUs required for the task, bandwidth thresholds, and latency constraints) and global resource status information (such as the number of remaining GPUs on each edge node, available bandwidth of the link, and energy consumption thresholds). Based on these inputs, the system can determine the gap between resource supply and demand without waiting for specific policy allocation results.

[0046] Secondly, when multiple tasks request the same type of limited resources, the system can abstract potential conflict relationships. For example, if three tasks request a total of 5 GPUs, but an edge node only has 3 GPUs remaining, then theoretically at least 2 of the requests cannot be satisfied simultaneously on that node. This conflict can be calculated upfront without relying on subsequent strategy generation. Based on this, the system models tasks as nodes in a graph, and potential conflicts between tasks due to sharing the same resources as edges, attaching attributes such as conflict intensity, resource availability, and request volume to the edges. This forms a resource competition graph that can quantify task-to-task conflict relationships.

[0047] Furthermore, this scheme uses the resource contention graph as a priori input for policy generation and optimization, enabling large decision-making models and reinforcement learning agents to directly perceive the coupling relationships between tasks and the location of bottleneck resources when generating policies. This "pre-conflict modeling" effectively avoids blind searching and repeated corrections during the policy generation stage, improving policy convergence speed and global optimality.

[0048] In summary, the principle of resource contention graph construction lies in: modeling potential conflict relationships in advance based on the difference between user demand and resource supply, and formally representing them through a graph structure, providing structured prior knowledge for subsequent strategy generation and optimization. Its innovation lies in shifting resource conflict identification from "post-strategy judgment" to "pre-strategy modeling," significantly enhancing the guidance and robustness of the strategy generation process.

[0049] Step S300: Construct a Prompt sequence based on all the task information and resource contention graph, input the Prompt sequence into a preset large decision model, and the large decision model outputs an initial strategy; In the specific implementation process, the large decision-making model adopts the LDM model. During the initial strategy construction process, LDM employs an improved Transformer architecture. The input sequence is divided into different modal subsequences (business requirement sequence, graph feature sequence, and historical fragment sequence).

[0050] The model introduces a cross-modal attention mechanism during the encoding stage to enable feature interaction between different modalities.

[0051] The model utilizes a multi-head attention mechanism during the encoding process to capture the coupling relationship between business needs and resource conflicts. For example, latency-sensitive tasks receive significant attention in high-load prediction environments.

[0052] Precedence and precedence constraints between tasks are maintained through positional encoding and task dependency encoding.

[0053] During the decoding phase, the model does not generate natural language, but instead directly outputs a structured policy template.

[0054] The output format is JSON / DSL, and it contains the following fields: placement: The mapping between tasks and resource nodes, such as {TaskA: Edge1-GPU2}; allocation: Resource quotas and scheduling priorities for each task, for example, {TaskA: 2GPU, Priority=High}; routing: Link and path selection, for example {TaskA→ServerX:Path[Edge1→Core→ServerX]}; fallback: a mechanism for reverting to previous behavior and migration rules, such as {if GPU > 95% → migrate to Edge2}.

[0055] After generating the initial policy, the component will automatically call the constraint check module to verify whether the draft violates hard constraints (such as insufficient resource capacity, unreachable links, and policy loops).

[0056] If a conflict is detected, LDM will enter a self-correction round, regenerating or adjusting some policy fields based on error feedback.

[0057] Output: A structured strategy draft, including strategies such as task deployment location, resource quota, routing scheme and fallback logic, which can be directly used by subsequent reinforcement learning evolution components.

[0058] In this approach, the initial policy proposal serves as the initial policy input for the reinforcement learning agent, providing a highly feasible and structured starting point for decision-making. Traditional reinforcement learning often requires starting with random policies, leading to slow early convergence and numerous ineffective actions. However, by introducing policy drafts generated by LDM, the agent can start from a near-feasible solution, significantly shortening the cold start training period and reducing resource waste and simulation costs caused by ineffective exploration. The innovation of this approach lies in combining generative decision-making models with reinforcement learning to form a closed-loop "generation-evolution" process, retaining the global search capability of LDM while leveraging the local optimization advantages of reinforcement learning.

[0059] Step S400: Construct a prediction vector based on the task request time point in the task information, and output the task prediction result using a time-series prediction model; In the specific implementation process, during the processing steps of outputting task prediction results using a time-series prediction model, Input historical resource load curves and task arrival rate logs.

[0060] The time-series forecasting model is used to predict the number and distribution of tasks arriving at each processing node within a future time window.

[0061] The resource utilization of each processing node within a future time window is predicted using a time-series forecasting model.

[0062] The predicted task arrival trends and resource load trends are embedded into the state vector for subsequent strategy generation.

[0063] Output the task arrival prediction and future load curve for each processing node.

[0064] The above scheme uses task arrival prediction and future load curves as important components of the state space, providing the agent in the reinforcement learning model with forward-looking information on future resource load and task arrival. Traditional reinforcement learning often makes decisions based on the current state, making it difficult to respond promptly to future load peaks or sudden task surges. By embedding the prediction curve and task arrival distribution into the state vector, the agent can consider potential bottlenecks and congestion in advance during policy evolution, thereby generating a forward-looking and robust scheduling policy. The innovation of this approach lies in jointly modeling time series prediction results with real-time states, enabling the policy to not only be optimal for the current situation but also take into account future evolution trends, thus improving adaptability in dynamic environments.

[0065] Step S500: Input the resource competition graph into a preset graph neural network model, and the graph neural network model outputs a graph embedding feature vector; In the implementation process, a graph neural network is used to compute the conflict graph, extracting node embeddings to represent task attributes, extracting edge embeddings to represent conflict intensity, and generating a global embedding to characterize bottleneck resources. Based on the node, edge, and global embedding vectors, task coupling features and conflict paths are derived for subsequent policy optimization. Output: Graph embedding feature vector.

[0066] Using the above approach, graph features extracted by the graph neural network are used to characterize the coupling relationships and resource conflict paths between tasks, and are embedded into the state space of reinforcement learning. Traditional reinforcement learning typically relies on vectorized state descriptions, which struggles to capture the complex dependency structures and resource competition relationships between tasks, easily leading to one-sided or locally optimal scheduling strategies. By introducing graph embedding vectors, the agent can identify strong couplings and bottleneck resource locations between tasks, thus enabling more targeted node selection, path selection, and resource allocation during action selection. The innovation of this approach lies in introducing graph structure awareness, giving reinforcement learning global topological awareness and dependency modeling capabilities, effectively improving scheduling accuracy and resource utilization in complex multi-task scenarios.

[0067] Step S600: Based on the initial policy, task prediction results, and graph embedding feature vectors, a reinforcement learning algorithm is used to output the final policy.

[0068] This scheme, employing a large decision model and reinforcement learning algorithms as its core, is responsible for generating and optimizing network service strategies. First, the fused business requirements and resource contention graph are uniformly encoded. A prompt sequence is constructed by combining multimodal information such as resource status, constraints, and historical strategy experience. This drives the LDM to generate a structured strategy draft covering task deployment, resource allocation, and path selection, and its feasibility is verified. Subsequently, a temporal prediction model (such as Informer or LSTM) is introduced to predict the changing trends of network resource load. A graph neural network (GNN) is then used to extract task coupling features and conflicting paths from the resource contention graph. Finally, the system uses the strategy draft as the initial policy input for the reinforcement learning agent. The prediction results and graph structure features are embedded into the state space, and policy training and evolution are conducted through a "state-action-reward" mechanism. This generates multiple candidate policy sets adaptable to different performance requirements and resource constraints, and ultimately selects a policy, balancing the flexibility of large-scale computing power scheduling with diverse service quality requirements.

[0069] This solution aggregates the structured resource requirements of multiple users and intelligently generates network service strategies such as task scheduling, routing planning, and resource allocation, which is the core link in service generation. Due to the high complexity and dynamism of business requirements, traditional methods that rely on static rule configuration or manual design are difficult to respond to changes in network status and user intent in real time. Furthermore, reinforcement learning algorithms, which have been widely used in recent years, generally face problems such as long training cycles, weak generalization ability, and insufficient adaptability to dynamic environments.

[0070] This solution's architecture incorporates a Large Decision Model (LDM) as the initial policy generator for network services. LDM is a model system designed for policy generation tasks, with core technologies including multimodal input fusion, adaptive policy generation, and structured output representation, aiming to support complex, high-dimensional, and unstructured intelligent decision-making scenarios. Compared to traditional models that only process single-type inputs, LDM can accept multimodal inputs such as natural language, images, and tables, possessing the ability to uniformly model heterogeneous information such as business requirements, environmental states, and external constraints. Based on this, LDM generates interpretable, executable, and optimizable policy outputs using a Transformer architecture, and integrates reinforcement learning and search optimization mechanisms to construct a closed-loop policy generation process of "generation—evaluation—optimization," thereby significantly improving the adaptability and robustness of the policies.

[0071] In some embodiments of the present invention, in the step of constructing an edge between two nodes if the resource requirements of any two service request tasks conflict, the resource types required by the service request tasks are read from the task information, and the resource types required by any two service request tasks are compared to see if they overlap. If they overlap, it is determined that the resource requirements of the two service request tasks conflict.

[0072] In some embodiments of the present invention, if the resource requirements of any two service request tasks conflict, the step of constructing an edge between the two nodes further includes constructing a label including identification information on the edge between the two nodes. The identification information includes the node in the service network requested by the service request task, the remaining resource quantity in the node in the service network requested by the service request task, the resource quantity requested by the service request task from the node in the service network, and the number of resources contested.

[0073] In some embodiments of the present invention, in the step of constructing a Prompt sequence based on all the task information and resource contention graph, prompt word engineering technology is used to construct a Prompt sequence from all the task information and resource contention graph.

[0074] In practice, the process of constructing the Prompt sequence includes: Conflict graph encoding: Using graph neural networks, the attributes of task nodes and the relationships of conflict edges in the resource competition graph are transformed into embedding vectors of nodes and edges.

[0075] Resource state coding: Transforms the time-series curves of resource utilization and bandwidth in the network into fixed-dimensional state vectors through a time-series coding model.

[0076] Historical experience retrieval and embedding: Extract corresponding scenarios from the strategy library and transform their deployment results and performance metrics into vectorized experience tips.

[0077] Prompt concatenation: Concatenates the embedding vector, state vector, and vectorized experience tips into a unified Prompt sequence in a predefined order.

[0078] In some embodiments of the present invention, in the step of outputting the final policy using a reinforcement learning algorithm based on an initial policy, task prediction results, and graph embedding feature vectors, the initial actions of the reinforcement learning algorithm are constructed based on the initial policy, and the initial state of the reinforcement learning algorithm is constructed based on the task prediction results and graph embedding feature vectors.

[0079] In some embodiments of the present invention, in the step of using a reinforcement learning algorithm to output the final policy, the reinforcement learning algorithm outputs a policy in each round of processing; During each round of processing, the updated state is determined based on the initial state and the selected action of each round. Based on the updated state, the business goal achievement indicators, cost items, energy consumption items, constraints, penalties, and stability items are determined. Based on the business goal achievement indicators, cost items, energy consumption items, constraints, penalties, and stability, the reward value is determined.

[0080] In some embodiments of the present invention, the reward value is calculated using the following formula in the step of determining the reward value based on business target achievement indicators, cost items, energy consumption items, constraints, penalties, and stability: Reward value=α*S-β*C-γ*E-λ*V-μ*M; Where S represents the business objective achievement index, α represents the weight corresponding to the business objective achievement index, C represents the cost item, β represents the weight corresponding to the cost item, E represents the energy consumption item, γ represents the weight corresponding to the energy consumption item, V represents the constraint penalty, λ represents the weight corresponding to the constraint penalty, M represents the stability item, and μ represents the weight corresponding to the stability item.

[0081] In some embodiments of the present invention, in the step of using a reinforcement learning algorithm to output the final strategy, the strategies output by the reinforcement learning algorithm in all rounds of processing are collected as an initial strategy draft, sorted according to the business objective achievement index of each strategy, and the final strategy is selected based on the sorting result.

[0082] In the process of reinforcement learning algorithms, a state space is defined, which includes task features, resource states, and prediction results.

[0083] Task characteristics provide a vectorized description for each service request task, including at least the business category, priority, deadline or expected latency, target packet loss rate or reliability threshold, estimated computing power requirements such as the number of GPU frames or video memory requirements, average bit rate or burstiness, deployment affinity and anti-affinity, allocated location and queue location, and runtime stage index. For tasks consisting of multiple stages, additional topological codes for stage dependencies, such as DAG order index or adjacency digest vector, are added to maintain pre- and post-constraints.

[0084] Resource status provides the current capacity and occupancy vectors, as well as steady-state and transient metrics, for each processing node and link. Processing nodes include at least the number of available GPUs and video memory, remaining CPU and memory, container slots, average queuing latency and drop rate, node energy consumption and carbon intensity, unit cost and cross-domain cost, and failure rate or fault rate. Links include at least available bandwidth, propagation and queuing latency, packet loss rate, jitter, path-level compliance restrictions, and costs.

[0085] The prediction results, the trend prediction component provides the number and distribution of tasks arriving within the future time window, as well as the resource utilization curves of each node and link, and gives the interval or confidence level; to ensure the Markov property approximation, the state includes the compressed history of the most recent steps, such as sliding window statistics, and time position encoding, such as relative time scale, so that the agent has memory in a partially observable environment.

[0086] Graph Structure Summary. The conflict graph between tasks and resources is processed by a graph neural network, and then node embedding vectors, edge embedding vectors, and global embeddings are added to characterize the coupling strength and bottleneck locations, enabling the agent to have structure awareness when making action selections.

[0087] State organization and specification. The above elements, after normalization and embedding, constitute a hierarchical state tensor, organized into three subspaces: task dimension, resource dimension, and global dimension. Unusable or restricted objects are marked with a mask matrix to ensure dimensional stability and online inference efficiency.

[0088] Define actions such as node selection, path selection, and resource quota adjustment strategies.

[0089] Node selection: The agent selects the target node for deployment or migration for the specified task from the set of candidate processing nodes, and determines the location where the task's computing power is carried.

[0090] Path selection: For task traffic transmission, the agent selects one or more feasible paths to carry the data, and determines the data forwarding method and link occupancy.

[0091] Resource quota adjustment: Under the premise that the task and target node are bound, the agent selects the share of computing resources to be allocated, such as the number of GPUs, the number of CPU cores or the bandwidth ratio.

[0092] The action output is a set of decision triplets, typically in the form of {task identifier, node selection / path selection, quota or scaling parameter}. This result is fed back to the reward function after execution in the simulation environment.

[0093] Define the reward as: Business SLA satisfaction rate - cost - energy consumption.

[0094] Basic target items. Define the business target achievement index S at the current decision step, including negative values ​​of task-level target achievement rate or latency exceedance rate, throughput achievement, etc.; define the cost item C, including computing resource cost, link cost and cross-domain or backhaul cost; define the energy consumption item E, including node energy consumption or unit service energy consumption, and weighted carbon intensity if necessary.

[0095] Constraints and Stability Items. A large penalty V is imposed for violations of hard constraints such as capacity, compliance, and affinity; a smoothing penalty M is imposed for service disturbances introduced by migration, frequent switching, and drain / rebuild, in order to suppress policy jitter and deployment instability.

[0096] Reward function and weights. Calculated using the following formula: Reward value=α*S-β*C-γ*E-λ*V-μ*M; Wherein, S represents the business objective achievement index, α represents the weight corresponding to the business objective achievement index, C represents the cost item, β represents the weight corresponding to the cost item, E represents the energy consumption item, γ represents the weight corresponding to the energy consumption item, V represents the constraint penalty, λ represents the weight corresponding to the constraint penalty, M represents the stability item, and μ represents the weight corresponding to the stability item. The weight vector can be adjusted online according to tenant profiles or policy preferences, and supports conditional weights, such as increasing α and decreasing β and γ during emergency periods. To alleviate sparse rewards, potential energy shaping can be introduced for terminal indicators or task completion rewards can be densified according to progress. To improve robustness, quantile or conditional value at risk targets can be adopted, and the weight of the worst-case scenario can be increased.

[0097] The overall goal of time-domain accumulation is to maximize the discount return, and an appropriate discount factor is used to balance short-term delay control with long-term energy efficiency and cost benefits.

[0098] Iteratively train the agent and gradually evolve the strategy.

[0099] Initialization and behavioral priors: The structured policy draft generated by LDM is used as the initial policy for behavioral priors or imitation learning, so that the agent starts from a highly feasible starting point and reduces early ineffective exploration; historical operation logs and twin simulation playback are used to construct offline trajectories for pre-training.

[0100] In the environment and interaction, each decision step is based on the current state and action mask, and the policy network outputs action parameters. After the environment executes the action, it returns a new state, an immediate reward, and optional constraint flags. Actions that cannot be performed are directly discarded and recorded as invalid samples.

[0101] In this implementation, the learning algorithm and stabilization methods preferably use policy gradient methods with constraint processing, such as proximal policy optimization or actor critic with Lagrange multipliers, combined with entropy regularization and parameter noise enhancement exploration. Target network, gradient pruning and normalization are used to improve training stability, and multi-head or conditional policies are used in multi-task and multi-domain scenarios to share underlying features.

[0102] Action masking and feasibility rollback are implemented. Safety masks are used during both the training and inference phases to avoid selecting unfeasible actions, and a quick feasibility check is performed again before execution. If the check fails, a local rollback or suboptimal alternative action is triggered to ensure uninterrupted online execution.

[0103] Curriculum and randomization, through curriculum-based progressively challenging training from small-scale single-domain training to large-scale cross-domain training, and using domain randomization based on demand distribution, arrival rate, link perturbation, and fault injection, enable the policy to have stronger generalization and robustness.

[0104] Model selection and candidate set generation: Periodically evaluate the policy on the validation set and retain snapshots of several policies with the best performance or policy variants with different weight settings to form a candidate policy set for subsequent multi-objective ranking.

[0105] Online fine-tuning and drift protection: After deployment, online fine-tuning is performed at a low learning rate limited by a safety threshold, and key indicator drift is monitored; when distribution drift or performance degradation is detected, it automatically rolls back to the most recent stable snapshot; and an initial policy draft is output.

[0106] In the step of selecting the final strategy, the optimal solution is selected from the set of candidate strategies and transformed into a standardized, executable strategy for use by the twin verification and service orchestration module.

[0107] The inputs are the initial strategy draft, predicted performance metrics, and business target weights.

[0108] Calculate the SLA compliance rate, overall throughput, energy consumption level, service cost, and latency jitter of the candidate strategies in the initial strategy draft under given business requirements.

[0109] Indicators with different dimensions and value ranges are normalized to a unified interval, enabling weighted comparison of indicators within the same evaluation system. A strategy performance scoring table is output.

[0110] The input scoring table is linearly weighted, and different performance indicators are summed according to the weight of business objectives to obtain the comprehensive score of the candidate strategies.

[0111] Perform Bayesian ranking, introduce uncertainty modeling, and rank the scores probabilistically to reflect the differences in credibility of different strategies.

[0112] Perform Pareto screening to preserve non-dominated solutions in the multidimensional objective space, ensuring that the output candidate policies outperform other policies in at least one dimension. Output a ranked list of policies, including the highest-scoring solution and several alternatives from the Pareto front.

[0113] Input the optimal strategy and convert the selected strategy into a unified DSL (Domain-Specific Language) format to ensure that the strategy is parseable and cross-platform portable.

[0114] Embed a fallback mechanism in the DSL description, such as automatically triggering task migration or resource reallocation when GPU utilization exceeds 95%, to enhance the robustness of the strategy.

[0115] Output a strategy description file conforming to the twin simulation interface standard, enabling it to be directly called, executed, and verified by the simulation module. Output a standardized strategy file for the final strategy.

[0116] In some embodiments of the present invention, in the step of receiving a service request task and parsing it to obtain task information, the service request task is parsed using a large language model to obtain the task information; In practical implementation, the large language model serves as the intelligent semantic mapping entry point between business requirements and network resources, responsible for accurately translating user natural language text into structured resource requirements. This module consists of three components: intent parsing, requirement translation, and configuration generation. It integrates key technologies such as natural language processing, knowledge graphs, and large language models, enabling text semantic parsing, business requirement reasoning, and resource configuration generation. This ensures a deep understanding and accurate mapping of user intent and provides structured resource requirements for the policy generation module. It is the logical starting point for achieving efficient resource allocation and intelligent service generation in this solution's architecture.

[0117] The method further includes verifying the final strategy in a twin environment using a twin mapping module; The twin mapping module provides policy verification and simulation optimization capabilities by constructing a virtual mirror of the real network. Composed of three components—twin modeling, simulation optimization, and data fusion—it leverages key technologies such as digital twins, fault simulation, and anomaly detection to effectively verify and optimize the generative network services output by the policy generation module within a twin environment. This provides executable policies for the service orchestration module, forming a closed-loop mechanism from twin construction and simulation verification to policy optimization. This module effectively reduces the risks of actual network operation and is a crucial guarantee for translating theoretical policies into practical deployments within this solution architecture.

[0118] The method also includes configuring the network according to the final policy via a service orchestration module.

[0119] In the specific implementation process, the intent parsing module parses the structured requirements output by the service request task through the large language model, such as the number of GPUs, latency requirements, and bandwidth requirements; the intent parsing module obtains the current network resource status, the number of remaining GPUs at the edge nodes, the available bandwidth of the link, the energy consumption threshold, and historical scheduling records (past task allocation and conflict situations).

[0120] The large language model employs bidirectional encoding models such as BERT and RoBERTa, autoregressive generative models such as GPT and OPT, and structural enhancement models such as ERNIE and T5.

[0121] For example, Standard format User A entered "Low-latency GPU inference", and User B entered "High-definition video processing".

[0122] Operation details: Units are normalized (e.g., bandwidth is standardized to Mbps, and latency is standardized to ms).

[0123] Complete the default fields (e.g., if no energy consumption requirements are specified → use the system default value).

[0124] Output: A uniform JSON representation, such as {task:T1, GPU:2, latency<20ms}.

[0125] Using the above approach, combined with a pre-built business knowledge graph and requirement mapping rules, a comprehensive and standardized expression of business requirements is generated. The system first infers the corresponding performance and functional requirement sets based on the identified business types and scenario information, achieving semantic expansion of implicit or omitted information in the original user intent. Then, it invokes the requirement mapping rules, combining a semantic matching mechanism and a numerical conversion table, to translate unstructured, ambiguous, or non-quantitative descriptions into formally standardized and measurable business requirements. For example, in the "high-definition video conferencing" scenario, the system can automatically complete its requirement dimensions in areas such as bandwidth and load balancing, translating "no lag" into "network latency <100ms" and "thousands of people online" into "total bandwidth ≥1Gbps".

[0126] Example: During the requirement aggregation phase, three users submitted video inference tasks: Task T1 requested 2 GPUs with a latency requirement of less than 30ms; Task T2 requested 2 GPUs with a latency requirement of less than 50ms; and Task T3 requested 1 GPU with a latency requirement of less than 100ms. Currently, edge node E1 has only 3 GPUs remaining. The system first converts the requirements into JSON format using a format unification component; then, it uses a semantic alignment component to map "low latency" to formal constraints such as less than 30ms; finally, a conflict modeling component constructs a task-resource conflict graph, marking GPU conflict intensity as 1. The output is a unified requirement set and a task-resource conflict graph.

[0127] In the policy generation phase, the system first constructs a prompt sequence, encodes the conflict graph into task embeddings and conflict edge embeddings, encodes the GPU utilization curve of E1 into a temporal state vector, retrieves similar deployment experience from the historical policy library and quantizes it, and finally concatenates them into a unified prompt sequence input. Subsequently, based on the Transformer, the input sequence is processed to generate a structured policy draft: T1 is deployed on E1, T2 is deployed on E1, and T3 is deployed in the cloud, along with the path and fallback conditions. After a feasibility check, the draft enters a correction round because the total requests on E1 exceed the remaining capacity, and is adjusted to deploy T1 on E1, T2 on E2, and T3 in the cloud. The prediction model further outputs that the GPU utilization of E1 will rise to 90% in the next fifteen minutes and predicts that two new tasks will arrive. This prediction result is embedded in the state vector to enhance the foresight of the policy. The GNN calculates the conflict graph and extracts tasks T1 and T2 that have high coupling and share the bottleneck resource E1. The reinforcement learning evolution component uses the revised draft as the initial policy and iteratively optimizes it based on the state-action-reward mechanism to generate three candidate policies: Option 1 is to keep T1 in E1, partially migrate T2 to E2, and execute T3 in the cloud; Option 2 is to distribute T1 and T2 to E1 and E2, and execute T3 in the cloud, selecting low-latency paths and high-bandwidth paths respectively; Option 3 is to have T1 in E1, T2 in E2, and T3 in the cloud, and set expansion clauses for T1 to improve robustness to burst loads.

[0128] The above-mentioned solution, with LLM technology at its core, is responsible for automatically transforming multi-dimensional business requirements into structured resource configuration schemes. The system first constructs a specific Prompt sequence based on business type, requirement parameters, and network status, clarifying the format, target dimensions, and output constraints of the generated tasks. Then, it uses a Transformer structure based on a self-attention mechanism to model the input information and generate directly executable resource configuration parameters. For example, in a high-definition video conferencing scenario, the system can combine specific latency and bandwidth requirements with the current network status to generate a configuration scheme such as "2 edge GPU nodes, 4 cloud GPU nodes, 500GB edge cache, and 7Gbps network bandwidth," providing a structured resource view to support subsequent global resource scheduling and service orchestration.

[0129] Overall, this technology is primarily based on LLM (Limited Language Management), achieving efficient transformation of user natural language input into directly invoked structured service requests through refined Prompt engineering design. Compared to traditional parsing methods that rely on rule matching or keyword extraction, LLM outperforms in semantic representation and generalization capabilities, significantly improving the accuracy of intent parsing in diverse, ambiguous, and implicit expression scenarios. This provides crucial support for SGN (Single-Native GN Service) to accurately parse and intelligently generate services under complex requirements.

[0130] During the strategy selection phase, the system first comprehensively evaluates the SLA compliance rate, throughput, energy consumption, cost, and latency jitter of the three candidate strategies and normalizes the results. For example, Scheme 2 has an SLA compliance rate of 99.5%, the best average latency, and moderate energy consumption and cost. The system then performs multi-strategy ranking, first obtaining a comprehensive score through linear weighting to confirm that Scheme 2 has the highest score, then using Bayesian ranking to further confirm the reliability of its SLA metrics, and finally using Pareto filtering to retain Scheme 2 and Scheme 1 as non-dominated solutions. Finally, the system converts Scheme 2 into a DSL format and embeds a fallback mechanism, such as automatically migrating the task to E3 when E1 GPU utilization exceeds 95%, and outputting a standardized strategy file that can be directly called by the twin simulation.

[0131] Finally, the system outputs the selected optimal strategy and delivers it to the twin simulation module for verification and testing. Verification results show that this strategy can meet the latency requirements of three types of video inference tasks while keeping resource utilization within a safe range, maintaining a high SLA compliance rate and low energy consumption.

[0132] In summary, with the rapid evolution of GAI technology, large-scale models have acquired capabilities such as complex semantic understanding, intent recognition, and knowledge reasoning, significantly improving the intelligence level of automated decision-making and policy generation. Simultaneously, the convergence of computing and networks is accelerating its evolution from the traditional "connection-scheduling" stage to the "intelligence-autonomy" stage, exhibiting new trends such as increased demand for multi-domain computing and network resource collaboration and the prominent self-configuration capabilities of service-oriented resources. Based on the intelligent generation capabilities of GAI technology and the collaborative orchestration system of the computing and network convergence concept, this invention proposes a new paradigm for service generation networks, aiming to achieve a paradigm shift in network services from manual intervention to intelligent generation, and construct a new intelligent service system driven by cognitive reasoning.

[0133] Specifically, this solution focuses on intelligent perception, policy generation, and collaborative scheduling in the network service generation process, possessing the following four core characteristics: 1. Intent-driven recognition: Based on user natural language input, it identifies their service intent and performance requirements, mapping them to executable resource needs; 2. Intelligent service generation: Adaptively generates personalized network service policies to meet dynamic network conditions and task demands, achieving on-demand scheduling and rapid response; 3. Global resource collaboration: Integrates multi-domain computing network resource status information to achieve integrated cloud-edge-device resource optimization and collaborative management; 4. Autonomous evolution and optimization: Through closed-loop feedback and continuous learning mechanisms, it continuously improves the effectiveness and adaptability of service policies.

[0134] The integration of these capabilities endows this solution with a higher level of intelligent autonomy and service adaptation, marking the evolution of the next-generation network architecture from "functional definition" to "integrated semantic-driven and policy generation." This provides key support for complex and diverse business needs and dynamically changing network environments, thereby solving the bottlenecks of computing power networks in unified resource abstraction, collaborative management, and real-time scheduling. It also makes up for the shortcomings of service-customized networks in intelligent semantic understanding, flexible orchestration, and closed-loop management throughout the entire lifecycle, significantly improving the network's efficient carrying capacity and differentiated guarantee capabilities in multiple scenarios, multiple objectives, and multiple spatiotemporal scales.

[0135] The beneficial effects of this plan include: (1) Breakthrough in unified abstraction and collaborative scheduling of resources While existing computing networks have achieved interconnection and scheduling of cross-domain computing power, they still face efficiency bottlenecks in the collaboration of multiple types of computing power (CPU, GPU, FPGA, edge nodes, etc.) due to the high heterogeneity of resources and insufficient abstraction standards. This invention introduces a GAI-driven semantic parsing and resource modeling mechanism to directly transform user business intent into standardized resource requirement descriptions, unifying resource abstraction from the source. Based on this, combined with multi-agent reasoning and multi-objective optimization algorithms, dynamic resource collaboration across domains and levels is achieved, thereby significantly reducing scheduling latency and improving resource utilization.

[0136] (2) Improvement in the intelligence of service generation and orchestration While service-customized networks can orchestrate services based on preset templates, they largely rely on static rules and structured logic, lacking adaptive capabilities. This invention, through an integrated closed loop of "perception-reasoning-decision-optimization," combines GAI, LLM, and reinforcement learning to achieve real-time business intent parsing and policy generation. Furthermore, it utilizes digital twins for simulation verification and policy iteration, enabling service orchestration to automatically adjust according to network conditions and business changes, avoiding performance degradation caused by rigid rules in traditional methods.

[0137] (3) Improvement of closed-loop management throughout the entire life cycle CPN and SCN primarily focus on single aspects of resource allocation or service deployment, lacking systematic support for service runtime status, anomaly detection, and continuous optimization. This invention embeds an execution status feedback mechanism throughout the entire service generation process, such as... Figure 4 As shown, the model is fine-tuned and transfer learning is performed based on feedback data to form a continuously evolving intelligent service generation system, thereby achieving full autonomous management of the service lifecycle.

[0138] (4) Enhanced capabilities in multidimensional optimization and adaptation to complex scenarios CPN optimization objectives are mostly focused on resource utilization and latency, while SCN focuses more on service SLA assurance. Both have limited capabilities in comprehensively balancing multiple objectives. The strategy generation module of this invention performs interpretable optimization among multiple objectives such as latency, bandwidth, reliability, energy efficiency, and cost, adapting to complex business scenarios with multiple modalities, objectives, and spatiotemporal scales (such as vehicle-road cooperation, industrial internet, and AIGC tasks in intelligent computing centers), ensuring optimal service operation under different constraints.

[0139] This invention also provides a service-oriented network scheduling system, which includes a computer device, a processor, and a memory. The memory stores computer instructions, and the processor executes the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps of the method described above.

[0140] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned service-oriented network scheduling method. The computer-readable storage medium may be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0141] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0142] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0143] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0144] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A network scheduling method for service generation, characterized by, The method is based on a service network comprising a plurality of processing nodes, and steps of the method comprise: receiving service request tasks and parsing to obtain task information, the task information comprising a request node and resource requirements of the tasks; constructing a resource competition graph based on the resource requirements of each service request task, in which each service request task is taken as a node, and an edge is constructed between two nodes if the resource requirements of the two service request tasks conflict; constructing a Prompt sequence based on all the task information and the resource competition graph, inputting the Prompt sequence into a preset macro decision model, and outputting an initial strategy by the macro decision model; constructing a prediction vector based on a task request time point in the task information, and outputting a task prediction result by a time series prediction model; inputting the resource competition graph into a preset graph neural network model, and outputting a graph embedding feature vector by the graph neural network model; outputting a final strategy by a reinforcement learning algorithm based on the initial strategy, the task prediction result, and the graph embedding feature vector.

2. The service-oriented network scheduling method of claim 1, wherein, In the step of constructing an edge between two nodes if the resource requirements of the two service request tasks conflict, the resource types required by the service request tasks are read from the task information, and it is determined that the resource requirements of the two service request tasks conflict if the resource types required by the two service request tasks coincide.

3. The service-oriented network scheduling method of claim 2, wherein, In the step of constructing an edge between two nodes if the resource requirements of the two service request tasks conflict, a label comprising identification information is constructed on the edge between the two nodes, the identification information comprising a node in the service network requested by the service request task, a remaining resource quantity in the node in the service network requested by the service request task, a resource quantity requested by the service request task to the node in the service network, and a competing resource quantity.

4. The service-oriented network scheduling method of claim 1, wherein, In the step of constructing a Prompt sequence based on all the task information and the resource competition graph, a Prompt sequence is constructed from all the task information and the resource competition graph by using a prompt word engineering technology.

5. The service-oriented network scheduling method of claim 1, wherein, In the step of outputting a final strategy by a reinforcement learning algorithm based on the initial strategy, the task prediction result, and the graph embedding feature vector, an initial action of the reinforcement learning algorithm is constructed based on the initial strategy, and an initial state of the reinforcement learning algorithm is constructed based on the task prediction result and the graph embedding feature vector.

6. The service-oriented network scheduling method of claim 1, wherein, In the step of outputting a final strategy by a reinforcement learning algorithm, the reinforcement learning algorithm outputs a strategy in each round of processing; In each round of processing, an updated state is determined based on an initial state of each round and a selected action, a business target achievement degree index, a cost item, an energy consumption item, a constraint penalty, and a stability item are determined based on the updated state, and a reward value is determined based on the business target achievement degree index, the cost item, the energy consumption item, the constraint penalty, and the stability.

7. The service-oriented network scheduling method of claim 6, wherein, In the step of determining a reward value based on a business target achievement degree index, a cost item, an energy consumption item, a constraint penalty, and a stability, the reward value is calculated by using the following formula: reward value = α*S - β*C - γ*E - λ*V - μ*M Wherein, S represents a business target achievement degree index, a represents a weight corresponding to the business target achievement degree index, C represents a cost item, β represents a weight corresponding to the cost item, E represents an energy consumption item, γ represents a weight corresponding to the energy consumption item, V represents a constraint penalty, λ represents a weight corresponding to the constraint penalty, M represents a stability item, and μ represents a weight corresponding to the stability item.

8. The service-oriented network scheduling method of any one of claims 1 to 7, wherein, In the step of outputting the final strategy by using the reinforcement learning algorithm, the strategies output by the reinforcement learning algorithm in the processing process of all rounds are collected as an initial strategy draft, the business target achievement degree indexes of each strategy are sorted, and the final strategy is selected based on the sorting result.

9. The service-oriented network scheduling method of claim 1, wherein, In the step of receiving the service request task and parsing to obtain the task information, the service request task is parsed by using a large language model to obtain the task information; The method further includes a step of performing twin environment verification on the final strategy by using a twin mapping module. The method further includes a step of performing network configuration according to the final strategy by using a service arrangement module.

10. A service-oriented generated network scheduling system, characterized by, The system includes a computer device, the computer device includes a processor and a memory, the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the system implements the steps of the method as claimed in any one of claims 1-9.

Images